Convex on X

Fans describe it as a whole backend in one place: reactive queries, automatic caching and invalidation, transactional mutations, type-safe code end to end, background jobs and file storage. One calls Convex components a "game changer", another says it is the backend to keep after going public, and a founder credits it for keeping a product free to run. Setup and the new CLI draw praise for being easy and fun.

The reservations are practical: the schema tab is "kinda useless for bigger projects", a company that has used Convex for two years says the learning curve was steep until the team understood how it works differently, and one user warns an app will scale badly or cost more than it should if you are not careful. One fan says a rival is shipping faster.

The unconvinced view is a single thread: the tradeoffs against running SQLite locally did not add up, and the local binary's performance and size are poor.

The benchmark post on AI models travelled furthest: one user found a new model about as knowledgeable about Convex as a cheap one at 43 times the cost, and others repeated the cost gibe; the same user later called another model "meh" at Convex. Bug reports are specific: false "document not found" warnings for records that are in the table, deploy overrides that push to the old URL and 404, and useQuery errors blanking a whole screen, which one user worked around with a fails-soft hook.

Support complaints are about waiting: four days with no reply, a stuck automated reply, an unanswered GitHub issue, and a note that Convex still does not work well in a monorepo with several instances. Two users lost access through GitHub, one after a login was flagged and one when OAuth stopped working. The cost worry is a warning to be careful about scaling, plus a user slightly concerned but staying for the convenience.

The SQLite thread is the only comparison with more than one voice: a user wanted Convex to make sense but found "the tradeoffs vs just running local sqlite didn't make a ton of sense", judged the local binary's performance and size rough, and allowed it might work as cloud sync. The person they were talking to agreed it may not suit the case and said they would tinker slowly.

The rest are one-offs. Neon and ElectricSQL come up in a sync-engines discussion where a user calls Convex the most popular sync engine right now, and someone argues Convex is the better pick over Neon and is asked why. A user fully moved to Cloudflare D1 keeps Convex as the option for anything "way more serious". One asks whether Convex could do practically anything Airtable does. One wants Convex's LLM inference weighed against OpenRouter or OpenAI under contract. One says they happily pay for Convex rather than run or vibecode it themselves, and a fan says a rival is outshipping it.

Getting-started questions are about the first hour: which of app, database and sign-in took longest to understand, whether reading the docs or copying the prompt gets you to a first working component faster, and what besides API keys felt confusing. One user says setup was "super easy to use and set up", and someone asks whether the team checks what Claude and Codex answer when asked about Convex, calling that part of the funnel.

Security questions follow a post about short-lived keys: are five-minute keys scoped, per agent or per task; can secrets be marked non-copyable; how is database security handled around these cases; and whether the managed runtime, still reporting Node 24.21.0, is patched against current vulnerabilities.

Beyond those: two people ask how to recover an account after GitHub login fails, two ask whether file storage works with other providers (one answers it is Convex-only for now, extendable through customMutations in convex-helpers), two ask what Convex does that others cannot, two ask about sandboxes and MCP (a team member replies both have been there for months), two discuss games, two ask about splitting a schema into components, and one asks about a size cap in mutations versus actions.

The sandbox ask comes with a comparison: one fan says a rival is "way outdoing you guys lately with their releases" and asks when sandboxes and auto-MCP are coming, and another calls sandboxes for agents the missing piece, thinking about tool isolation without a full VM per call. A team member replied that several sandboxes and an MCP server already exist and have for months, so the ask is partly about people not knowing.

The single asks are concrete. A user wants a Mumbai region because even hobby projects have to choose the US. A developer who builds on Convex notes native support for React and React Native but not Flutter. Someone asks when Windows. A user of the AI docs chat likes the per-message copy button and wants one for the whole conversation. One wants fast key revocation and usage alerts because short expiry is not enough when a key leaks. Two are component ideas: a databuddy component so .track() works beyond buttons, and sub-components that each serve their own admin UI.